mlboydaisuke/RAFT-small-OpticalFlow-ExecuTorch

raft_small — ExecuTorch

• Source: torchvision raft_small (C_T_V2, FlyingChairs + FlyingThings3D)
• License: BSD-3-Clause
• Input: [[1, 3, 384, 512], [1, 3, 384, 512]] — two RGB frames, each scaled to [-1,1], 384x512 (both dimensions must stay divisible by 8)
• Output: flow [1,2,384,512] in pixels: channel 0 is horizontal displacement from frame 1 to frame 2, channel 1 vertical. Refined over 12 iterations, which are baked into the graph — the intermediate iterates are not returned.

Variants

All variants take and return fp32 tensors — swap the .pte file, keep your app code.

build	file	size (MB)	parity vs fp32 eager (worst corr)	Mac median (ms)*
fp32	raft_small_xnnpack_fp32.pte	4.4	1.000000	169.6

*Mac arm64, single process, median of 10 — a reference point for relative cost only, not a device number (torch eager fp32 on the same machine: 135.2 ms).

Verification (executorch 1.4.0, torch 2.13.0)

Parity is measured against the fp32 eager model on real image input; corr is the correlation over all elements of each output tensor.

output	shape	max_abs_diff	corr
0	[1, 2, 384, 512]	4.792e-04	1.000000

XNNPACK delegate coverage (fp32): 62.9% (1208/1921 ops); ops left on the portable kernels: dim_order_ops._to_dim_order_copy.default x435, aten.split_with_sizes_copy.default x50, aten.grid_sampler_2d.default x48, aten.expand_copy.default x30, aten.arange.start_step x28, aten.lt.Scalar x24, aten.sub.Tensor x24, aten.where.self x24, aten._native_batch_norm_legit.no_stats x21, aten.alias_copy.default x12, aten.view_copy.default x6, aten.avg_pool2d.default x3, aten.cat.default x2, aten.unsqueeze_copy.default x2, aten.repeat.default x2, dim_order_ops._clone_dim_order.default x1, aten.sqrt.default x1

Conversion

torch.export -> to_edge_transform_and_lower(partitioner) -> .pte (conversion scripts: executorch-models)

Notes: No Core ML build: RAFT's correlation volume is a rank-6 tensor and Core ML caps at rank 5. No fp16 build either — this is a conv-only model, where XNNPACK serializes convolution weights as fp32 whatever the graph dtype, so fp16 would be the same size with extra casts.

---

More models in this format: ExecuTorch Model Zoo — 31 models, each with the recipe that produced it.

Want a different model on-device? Open a request — free, open weights only; the export and its measured numbers get published publicly.